Pattern Recognition Letters 109 (2018) 120–128

Contents lists available at ScienceDirect

Pattern Recognition Letters

journal homepage: www.elsevier.com/locate/patrec

Review on mining data from multiple data sources

Ruili Wang a, Wanting Ji a,∗, Mingzhe Liu b, Xun Wang c, Jian Weng d, Song Deng e,
Suying Gao f, Chang-an Yuan g
a
Institute of Natural and Mathematical Sciences, Massey University, Room 2.14, Mathematical Sciences Building, Auckland 0632, New Zealand
b
School of Network Security, Chengdu University of Technology, Chengdu 610059, China
c
School of Computer and Information Engineering, Zhejiang Gongshang University, Hangzhou 310018, China
d
College of Cyber Security, Jian University, Guangzhou 519632, China
e
School of Computer Science, Nanjing University of Posts and Telecommunications, Nanjing 210043, China
f
School of Economics and Management, Hebei University of Technology, Tianjin 300401, China
g
School of Computer and Information Engineering, Guangxi Teachers Education University, Nanning 530023, China

a r t i c l e i n f o a b s t r a c t

Article history: In this paper, we review recent progresses in the area of mining data from multiple data sources. The
Available online 31 January 2018 advancement of information communication technology has generated a large amount of data from dif-
ferent sources, which may be stored in different geological locations. Mining data from multiple data
Keywords:
sources to extract useful information is considered to be a very challenging task in the field of data min-
Multiple data source mining
Pattern analysis
ing, especially in the current big data era. The methods of mining multiple data sources can be divided
Data classification mainly into four groups: (i) pattern analysis, (ii) multiple data source classification, (iii) multiple data
Data clustering source clustering, and (iv) multiple data source fusion. The main purpose of this review is to systemat-
Data fusion ically explore the ideas behind current multiple data source mining methods and to consolidate recent
research results in this field.
© 2018 Elsevier B.V. All rights reserved.

1. Introduction
The advancement of information communication technology
has generated a large amount of data from different sources, which
may be stored in different geological locations. Each database may
have its own structure to store data. Mining multiple data sources
[1–3] distributed at different geological locations to discover use-
ful patterns are critical important for decision making. In particu-
lar, the Internet can be seen as a large, distributed data repository
consisting of a variety of data sources and formats, which can pro-
vide abundant information and knowledge.
Data from different sources may seem irrelevant to each other.
Once information generated from different sources is integrated,
new and useful knowledge may emerge. Here is an excellent ex-
ample of how an organization to utilize mining data from different
data sources to obtain profound information, which cannot obtain
from an individual source.
The Australian Taxation Office (ATO) mines data from differ-
ent data sources such as social media posts, private school records
and immigration data to detect tax cheats. Mining data from dif-
ferent data sources become a sophisticated tool to crackdown tax
∗
Corresponding author.
E-mail address: jwt@escience.cn (W. Ji).
cheats that yielded nearly $10 billion in 2016 [4]. For example, in
a normal Australian family, the husband has a business and re-
ported $80,0 0 0 of taxable income per year, putting him just in-
side the second-lowest tax bracket, and his wife reported earning
$60,0 0 0 per year. But the data collected from different data sources
revealed that the family had three children at private schools at
an estimated cost of $75,0 0 0 per year, while immigration records
and social media posts showed that the family had recently taken
five business-class flights and a holiday in a Canadian ski resort,
Whistler. It means their declared incomes did not match their
lifestyle. This prompted ATO to contact them to confirm if they
have unpaid taxes. From the above example, we can see that devel-
oping an effective data mining technique for mining from multiple
data sources to discover useful information is crucially important
for decision making.
However, how to efficiently mine quality information from mul-
tiple data sources is a challenging task for current research [5–9],
especially in the current big data era, because in real world appli-
cations, data stored in multiple places often have conflictions [10].
The conflictions include: (i) data name conflictions: (a) the same
object has different names in different data sources, or (b) two
different objects from different data sources may have the same
name; (ii) data format conflictions: the same object in different
data sources has different data types, domains, scales, and preci-

https://doi.org/10.1016/j.patrec.2018.01.013
0167-8655/© 2018 Elsevier B.V. All rights reserved.
 R. Wang et al. / Pattern Recognition Letters 109 (2018) 120–128
sions; (iii) data value confliction: the same object in different data
sources records different values; (iv) data sources confliction: dif-
ferent data sources have different database structures.
In order to overcome these conflictions, four effective ap-
proaches have been adopted to mine useful information and dis-
cover new knowledge from multiple data sources: (i) pattern anal-
ysis [11–14], which mining useful patterns and information from
one data source or several data sources in accordance with chang-
ing conditions constraints or relationships; (ii) multiple data source
classification, which labels data sources according to a certain stan-
dard, then classifies them by their labels; (iii) multiple data source
clustering, which clusters data sources according to their similar-
ities/distances; (iv) multiple data source fusion, which combines
data from multiple data sources to achieve higher accuracy and
more specific ratiocinations. Based on these three approaches, we
can mine useful information from multiple data sources to discover
new knowledge according to individual needs.
The main purpose of this review is to systematically overview
current multiple source data mining methods and to consolidate
recent research results in this field. In this paper, the pattern anal-
ysis approaches for multiple data sources are reviewed first, and
the approach of multiple data source classification and clustering
is reviewed after that. Then, mining multiple data sources using
data fusion is reviewed. The rest of this paper is organized as
follows. Section 2 shows typical methods of multiple data source
mining using pattern analysis, while Section 3 is about multiple
data source classification and clustering. Section 4 describes meth-
ods of multiple data source fusion. Finally, Section 5 provides con-
clusions and discussions.
2. Key methods for pattern analysis
In this section, we will describe different methods of pattern
analysis on multiple data sources. Pattern analysis on multiple data
sources is a process of mining valuable information or extract-
ing hidden predictive information from different databases. Pat-
terns existing in multiple data sources can be categorized as lo-
cal patterns and global patterns. A pattern identified from a mono-
database can be seen as a local pattern, while a global pattern has
to be determined based on all the obtained local patterns from
multiple data sources [15–17]. Correspondingly, two mainstream
methodologies of pattern analysis are developed, named local pat-
tern analysis and global pattern analysis, which will be discussed
as follows.
2.1. Local pattern analysis
Local patterns analysis is a process of identifying patterns from
a mono-database, which can efficiently extract knowledge from
the mono-database [18]. Based on previous studies [19–25], we
can summarize the local pattern analysis methods into three cate-
gories: association rule mining, sequential pattern mining, and oth-
ers.
2.1.1. Association rule mining
Association rule mining is a process of discovering the probabil-
ity of the co-occurrence of items in a database. Association rules
express the relationships between co-occurring items, which are
often used to analyze sales transactions. Association rule mining is
also known as market basket analysis. There are four parameters
used to describe an association rule: support, confidence, expected
confidence, and lift.
For example, a transaction database of a supermarket records
the number of customers on a given day is 10 0 0. Among them,
there are 100 customers bought both bread and milk, the support
of bread → milk (bread and milk are two itemsets, bread → milk
121
presents an association rule between bread and milk) is 10%. If 40%
of customers who bought bread will also buy milk, then the confi-
dence of bread → milk is 40%. If there are totally 200 customers
bought milk in the given day, then the expected confidence of
bread → milk is 20%. Thus, the lift of bread → milk is 40%/20% = 2.
While the support measures the importance of an association
rule, the confidence measures the accuracy of an association rule.
The support can reflect the representativeness of an association
rule. Although some association rules have high confidences, their
supports are very low, which means these association rules have
little chance of being practical so that they are also unimportant.
The expected confidence of an association rule can be seen as
the support of an itemset without the influence of other itemsets,
while the lift of an association rule is the ratio of confidence to
expected confidence, which is used to describe how one itemset
affects another itemset. In general, the lift of a useful association
rule should be greater than 1. Only when the confidence of an as-
sociation rule is greater than its expected confidence, it shows a
correlation between these two itemsets (i.e., one itemset has a pro-
moting effect on another itemset), otherwise, this association rule
is meaningless.
According to the definition above, any two itemsets in a trans-
action can have an association rule between them. In fact, users
are only interested in the association rules that satisfy a certain
support and confidence. Therefore, a support-confidence frame-
work is used to mine the correlations between itemsets in the lo-
cal databases. Two thresholds, the minimum support and the min-
imum confidence, specify the minimum requirements of a mean-
ingful association rule.
Association rule mining consists of two phases: (i) In the first
phase, all the high frequent itemsets are identified. (ii) In the sec-
ond phase, association rules are generated from the high frequent
itemsets. It has been extensively studied to identify the relation
existing in databases between itemsets [26].
A classical algorithm of association rule mining named Apri-
ori algorithm is presented by Agrawal and Srikant [27], which is
used for mining frequent itemsets and relevant association rules.
A key concept in the Apriori algorithm is the anti-monotonicity of
the support measure. It assumes that: (i) All subsets of a frequent
itemset must be frequent. (ii) For any infrequent itemset, all its su-
persets must be infrequent. The algorithm can be divided into two
phases. In the first phase, it applies minimum support to find all
the frequent sets with k items in a database. In the second phase, it
uses the self-join rule to find the frequent sets of k + 1 items with
the help of frequent k-itemsets. Repeat this process from k = 1 to
the point when we are unable to apply the self-join rule. The Apri-
ori algorithm can find the relationship between items in a database
and it is the first efficient and practical algorithm for association
rule mining. But it needs to generate a large number of candidate
itemsets and repeatedly scan the entire database.
A Frequent-Pattern tree structure called FP-tree is presented by
Han et al. in [28], which is constructed by (i) Scan a database to
get all the frequent itemsets and their respective support in the
database, sorting frequent items in frequent itemsets in a descend-
ing order of their support values. (ii) Create the root node and
scan the database again. Then select the frequent items and sort
them in the order of frequent itemsets. It is used for collecting
compacted and important information about frequent patterns. Ac-
cording to FP-tree, a corresponding method called frequent-pattern
growth algorithm is proposed for mining the complete set of fre-
quent patterns [28]. The proposed method can efficiently avoid the
process of candidate generation-and-test, and it can cut down the
time cost of frequent pattern mining. However, the algorithm may
have a challenge when encountering graph-based patterns or noise
samples.
122 R. Wang et al. / Pattern Recognition Letters 109 (2018) 120–128
Different from traditional association rule mining and inspired
by the paradigm named Ratio Rule (RR), which is used for ac-
quiring the quantitative association knowledge, Yan et al. [29] pro-
posed a method that uses a robust and adaptive one-pass method
to mine ratio rules from a database independently. Then all the ob-
tained ratio rules will be integrated into a probabilistic model. The
method is called Robust and Adaptive Ratio Rule (RARR). The main
notion of RARR is converting the ratio rule problem into the min-
imization problem of a non-negative energy function. The greatest
advantage of RARR is robustness in dealing with outliers while tra-
ditional multiple-source ratio rule mining algorithms cannot solve.
Besides, it is adaptive to mine the rules for a dynamic data source.
A least association rule means that the association rule consists
of the least item, which is very important and critical when it is
used to find the infrequent or abnormal itemsets. Abdullah et al.
[30] proposed a method called Critical Relative Support (CRS) to
mine the significant least association rules, which satisfies a cer-
tain predefined minimum critical threshold value. The experimen-
tal dataset is from the examination results of computer science
students from University Malaysia Terengganu. CRS has been veri-
fied efficiently for discovering the significant least association rules
and can substantially decrease the number of unwanted rules. But
the performance of CRS has only tested on one educational dataset,
which should be tested by a variety of different real datasets.
In general, association rule mining uses the support-confidence
framework to find the rules of users’ interests. However, the
support-confidence framework cannot reflect the semantic mea-
sure among the items. Semantic measure is characterized by utility
values, which are associated with transaction items typically. The
items that cannot be measured directly by a decision maker such
as stock, cost or profit, are called utility. For example, the confi-
dence of an association rule represents the percentage that shows
how frequently the rule occurs among all rules.
High utility itemsets mining can be seen as an extension of fre-
quent itemset mining [31], which includes three requirements: (i)
The item appears more than once. (ii) The item has its own profits.
(iii) The mining goal is to find the items which have high profits.
High utility itemsets mining will be more valuable than a single
support-confidence framework and can permit with more accurate
financial analysis and decisions.
A new utility-confidence framework based association rule min-
ing method is proposed by Sahoo et al. in [31], which is also the
first research on mining non-redundant association rules extracted
from the high utility itemsets. It includes two-step process: (i)
find all high utility itemsets and (ii) extract all valid rules. A com-
pressed representation for association rules is generated with the
help of high utility closed itemsets (HUCI) and their generators.
The compressed representation is named as high utility generic ba-
sis (HGB), which contains rules with minimal antecedent and max-
imal consequent. After that, in order to extract the non-redundant
rule set, HUCI and their generator are mined by the proposed al-
gorithm HUCI-Miner that identifies HUCI and associates the high
utility generators to the corresponding closed itemset. Moreover,
an inference algorithm is proposed to derive all valid association
rules under this framework. Experimental results show a signif-
icantly better achievement in compactness of the proposed rule
mining algorithms.
2.1.2. Sequential pattern mining
As indicated in [32], a sequence is an ordered list of itemsets.
For example, in customer transactions, an itemset corresponds to
a transaction, and a sequence corresponds to a list of transactions
sorted by transaction time. The goal of sequential patterns min-
ing is to identify the maximal sequences satisfied a pre-specified
minimum support, and the maximal sequence can be considered
as a sequential pattern, which is a very important pattern used for
identifying the sequential relationships between different itemsets.
There is no doubt that studying sequential patterns is always an
active area of data mining [32].
Agrawal and Srikant [32] first introduced the concept of se-
quential pattern mining and proposed two similar algorithms
based on the Apriori algorithm for mining sequential patterns,
called AprioriAll and AprioriSome. The only difference between
them is that AprioriAll counts all the large sequences including
non-maximal sequences while AprioriSome only counts the maxi-
mal sequences. Experimental results show that both AprioriAll and
AprioriSome have the excellent scale-up properties because they
scale linearly as the number of itemsets or transactions. However,
in the process of candidate sequences generation, when the num-
ber of candidate sequences is too large or the length of a can-
didate sequence is too long, the memory space required for both
proposed algorithms will become much huge.
A novel sequential pattern mining method, called PrefixSpan
(i.e., Prefix-projected Sequential pattern mining) is proposed by
Han et al. [33], which aims to reduce the cost of candidate sub-
sequence generation and the size of projected databases. The
proposed method only examines the prefix subsequences rather
than the whole sequences. Then only the postfix subsequences
of these prefix subsequences corresponding to will be projected
into the projected databases. In each projected database, the pro-
posed method only mines local frequent patterns to obtain sequen-
tial patterns. Moreover, in order to improve mining efficiency, two
kinds of database projections, named level-by-level projection and
bi-level projection, are explored in [33]. Besides, a main-memory-
based pseudo-projection technique is developed for saving the cost
of projection and speeding up computation processing.
By identifying potential trends in data, approximate sequence
pattern can summarizes and represents the local database effec-
tively. An alternative local pattern analysis method for sequential
patterns mining based on a local database of a multi-database is
proposed by Kum et al. [34]. In order to mine approximate se-
quential patterns, called consensus patterns, the proposed method
presented a novel algorithm, called ApproxMAP (for Approximate
Multiple Alignment Pattern mining). The processing of ApproxMAP
includes two phases. In the first phase, all sequences are clustered
according to their similarity/distance. In the second phase, con-
sensus patterns are mined from each cluster through ApproxMAP
directly. The experimental results show that ApproxMAP can ef-
ficiently summarize a local database and reduce the cost for the
following global mining.
An efficient sequential pattern mining method based on item
co-occurrences is proposed by Fournier-Viger et al. [35] for can-
didate sequences pruning. It proposed a data structure called
Co-occurrence MAP (CMAP) for storing co-occurrence informa-
tion. Then the CMAP is used to prune candidates in three verti-
cal algorithms (SPADE (Sequential PAttern Discovery using Equiv-
alence classes) [36], SPAM (Sequential PAttern Mining) [37] and
ClaSP (Closed Sequential Patterns algorithm) [38]). Extensive ex-
perimental results based on real-world datasets show that the co-
occurrence-based pruning proposed in [35] is effective and CMAP
can reduce a substantial amount of time for candidate sequence
selecting.
Recently, in order to deal with complex real-world data mining
problems, how to mining frequent patterns with periodic wildcard
gaps has gained much research interest. A frequent pattern with
periodic wildcard gaps is to add gap constraints to normal frequent
patterns. It requires that the user-specified gap constraints be sat-
isfied between any two of frequent patterns. The contents between
the gaps can be any value. Usually, it is a predetermined symbol,
which is referred to as a wildcard collectively. Frequent patterns
with periodic wildcard gaps can reflect the common characteristics
of items in a certain gap. User-specified patterns will be mined di-
 R. Wang et al. / Pattern Recognition Letters 109 (2018) 120–128
rectly without a large number of meaningless patterns generated
at the same time.
Previous methods usually use matrices or other linear data
structures for frequent patterns with periodic wildcard gaps min-
ing, which consume a large amount of memory and running time.
An Incomplete Nettree structure is proposed by Wu et al. [39],
which is used to calculate the number of supports of its super-
patterns in a one-way scan. After that, two novel algorithms,
named MAPB (Mining sequentiAl Pattern using incomplete Net-
tree with Breadth first search) and MAPD (Mining sequentiAl Pat-
tern using incomplete Nettree with Depth first search), are used to
solve the problem above with low memory requirements. More-
over, a heuristic algorithm called MAPBOK (MAPB for top-K) is
proposed to process the Top-K frequent patterns in each length
[39]. Experimental results based on a real-world biological dataset
demonstrate the superiority of the proposed algorithms in running
time and memory consumption.
2.1.3. Other pattern mining
Yan and Han [40] proposed a method called graph-based Sub-
structure pattern mining (gSpan) for frequent graph-based pattern
mining. A lexicographic order is constructed by gSpan firstly. Then
gSpan mines frequent substructure patterns without the process of
candidate generation by using depth-first search strategy. gSpan is
the first algorithm to apply depth-first search in the field of fre-
quent sub-graph mining. It also can be applied to mining large fre-
quent subgraphs.
For a multi-branch company, how to make decision based
on a branch’s patterns is an important problem. Suganthi and
Kamalakannan [41] proposed a new clustering method for ex-
act synthesized frequent items and a new method to deal with
exceptional patterns. The first one is used to provide detailed
information of the associations among different types of items in
a market. The second one is used to process the transportation
cost of a company, manpower and non-profitable items. Experi-
mental results show that when the items in a database are greatly
associated, the proposed two methods work well.
2.2. Global pattern analysis
Local pattern analysis is to obtain frequent patterns form local
databases. Then these obtained frequent patterns will be synthe-
sized for global pattern analysis. The key of global pattern analysis
is how to set up synthesizing models.
In the process of global pattern analysis, in addition to synthe-
sizing the support and confidence of an association rule, the fre-
quency of an association rule is also an important pattern. The
frequency of an association rule is the number of data sources
which contain this association rule. According to a threshold set-
ting, an association rule can be high-frequent, low-frequent, or nei-
ther high-frequent nor low-frequent in different data sources [45].
Heavy association rules, high-frequent association rules and ex-
ceptional association rules are three specific types of association
rules for synthesizing global patterns from local patterns in dif-
ferent data sources [45]. A heavy association rule is an associa-
tion rule with high-support and high-confidence in multiple data
sources. A high-frequent association rule is an association rule
with high extraction in multiple data sources. An exceptional as-
sociation rule is a heavy association rule but low-frequent. A high-
frequent association rule is not necessarily a heavy association
rule. Based on these definitions, two algorithms are proposed by
Adhikari and Rao [45]. The proposed methods synthesized heavy
association rules in multiple databases and notified whether a
heavy association rule is high-frequent or exceptional.
In practice, high-frequent association rules are important for
decision making. A high-frequent association rule means the rule
123
is supported by most of databases and has larger chance to be-
come useful rules in the union of all databases. A weighting model
is proposed by Wu and Zhang [19] for synthesizing high-frequent
association rules from multiple databases. In order to extract high-
frequent association rules from multiple databases, the proposed
method constructed a new association rule selection method to en-
hance the proposed weighting model. In the rule selection method,
high-frequent association rules will be taken as relevant rules,
while low-frequent rules will be taken as irrelevant rules. More-
over, when mining association rules from unknown data sources,
a relative synthesizing model is constructed to synthesize these
association rules by clustering of local association rules. However,
the proposed weighting model has a limitation, which needs to be
built on the databases that have the same type of data sources.
For example, if a company with several branches has three types
of businesses: super markets, banks and investment trusts, all data
sources need to be classified into three classes according to the
business types, then the data sources in each class can be synthe-
sized by using the proposed weighting model.
In order to overcome the limitation of the above proposed
weighting model, an improved weighting model is proposed by
Ramkumar and Srinivasan [43] for synthesizing high-frequency
rules. A data source weight is calculated based on its transaction
population (i.e., the number of itemsets) of this data source. Based
on the weights of data sources and the supports of local databases,
the proposed methods can calculate the global support and global
confidence. Experimental results show that the global support and
confidence calculated by the improved weighting method is much
close to the mono-mining (i.e., putting all databases into one big
database to mine patterns) results.
Global pattern analysis intends to find a synthesized rule from
local patterns in multiple databases to match with the mono-
mining results closely. Some models such as weighting model and
minimum supporting principle are based on linear combination.
However, when the extracted patterns are not satisfying a mini-
mum support threshold value, they cannot be used for the pat-
tern synthesizing process. In some cases, this requirement may
ignore some useful information/patterns that are meaningful for
a right decision making. Thus, an improved probabilistic model
based on the maximum entropy method is proposed for the syn-
thesizing process by Khiat et al. [44]. Both clique method (a grid-
based spatial clustering algorithm) and bucket elimination tech-
nique [44] are used for reducing the complexity of the maximum
entropy method. The experiments are based on the frequent item-
sets which are discovered by the Apriori algorithm, and the results
show the efficiency of the proposed synthesizing method [44].
Zhang et al. [42] propose a method for synthesizing local pat-
terns in multiple databases, which is especially suitable for finding
global exceptional patterns. It considers that different data sources
should have different minimal supports due to they are comprised
of different databases. Therefore, it adopts a method of comput-
ing a pattern’s importance by measuring the distance between a
pattern’s support and the corresponding database’s minimal sup-
port. However, the proposed method does not consider that dif-
ferent data sources may have different effects in the synthesizing
process, which may exist in reality [42].
Different from many other approaches that do not consider the
relationship between the data in one database, a nonlinear method
is proposed by Zhang et al. [20] to find the nonlinear relations
of data in a database. The proposed method is called Kernel Esti-
mation for Mining Global Patterns (KEMGP). The proposed method
uses the kernel estimation method (a mapping method used Gaus-
sian kernel) to capture the patterns of nonlinear relation when
synthesizing local patterns for global patterns. The KEMGP can bet-
ter characterize that different databases have different contribu-
tions to the global pattern.
124 R. Wang et al. / Pattern Recognition Letters 109 (2018) 120–128
2.3. Others
Besides, other techniques such as compressed sensing [46] and
sparse learning [47] also are widely used in pattern analysis. Since
a signal can be expressed in Fourier series or Taylor series, a new
input signal can be reconstructed by several known signals.
The compressed sensing theory [46] considers that, as long as
an input signal is compressible or sparse in a transform domain,
by using an observation matrix, which is not related to the trans-
form base, the transformed high-dimensional input signal can be
projected onto a low-dimensional space. Then the input signal can
be reconstructed by solving an optimization problem from a small
number of projections with a high probability.
The purpose of signal sparse representation is to represent an
input signal with as few known signals (i.e., atoms in a given over-
complete dictionary) as possible. A more concise representation of
the input signal will be obtained so that the information contained
in the input signal can be used in the further processing like com-
pression and encoding. In pattern analysis, the sparse representa-
tion is used as follows.
In the process of global and local learning, minimizing both
global and local learning cost functions is much helpful to utilize
both local and global discriminative information of a target do-
main [48]. Based on this premise, a novel Multi-source Adaptation
learning method with Global and Local Regularization (MAGLR) is
proposed by Tao et al. [48] for visual classification tasks by ex-
ploiting multi-source adaptation regularization joint kernel sparse
representation (ARJKSR). ARJKSR uses target and source domains to
reconstruct a dataset by sparse representation. Based on the mixed
Laplacian regularization (i.e., global sparsity preserving Laplacian
regularization and local learning Laplacian regularization) semi-
supervised learning diagram, MAGLR can obtain a label prediction
matrix for unlabeled instances from the target domain. The pro-
posed methods considered correlations among different source do-
mains and applies sparse representation to multiple source do-
mains including local and global regularizations simultaneously.
However, how to select the optimal parameters for the multi-
source domain adaptation settings has not been solved, which is
still an open problem.
A multiple measurement vector (MMV) model is used to learn
the properties of the distributions of the sources in the recovery
processing and improve the recovery of the sources in [49]. By
using a mixture of Gaussians prior, which is inspired by existing
Sparse Bayesian Learning approaches, the proposed method is used
for simultaneous sparse approximation. The proposed method has
been tested on a variety of synthetic, simulated and real data with
good performance. The proposed method assumes that there is no
prior knowledge about the sources besides the number of compo-
nents in the Mixture of Gaussians distribution. In some cases, im-
proved performance can be achieved by taking further knowledge
of the prior into account.
In practice, global patterns in multiple databases are synthe-
sized according to the independent characteristics of local pat-
terns. The pipelined feedback technique (PFT) [50] is one of the
most effective techniques in local pattern analysis (LPA) for this
task. Before PFT is proposed, the Single Database Mining Technique
(SDMT) was used to mine local databases in multiple databases.
However, SDMT has a limitation related to sizes of databases: (i)
when each of the local databases is small, SDMT can mine the
union of all databases; (ii) when at least one of the local databases
is large, SDMT can mine every local database, but fail to mine the
union of all local databases; (iii) when at least one of the local
databases is very large, SDMT fails to mine every local database
[50]. In PFT, once a pattern is extracted from a local database, then
it will be extracted from the remaining databases. PFT improves
synthesized patterns as well as local patterns analysis significantly.
In order to analyze the effect of database grouping on multi-
database mining, an approach of non-local pattern analysis (NLPA),
which combined database grouping algorithm and PFT, is proposed
in [51]. Non-local pattern analysis is more attractive than LPA
when the frequency of data mining is lower, which is helpful to
estimate a lesser number of patterns when synthesizing the global
pattern. Experimental results show the disadvantage of non-local
pattern analysis is that the pattern reported only from the last few
groups which may result in the experimental error worse. This is
due to the absence of a pattern reported which has been assumed
to exist. Thus, the definition of the depth of a pattern is proposed
to solve this drawback. It means a pattern will become useless if
its depth is low.
3. Key methods for data source classification and clustering
In real world applications, a large amount of data from differ-
ent sources normally is stored in different geological locations. In
order to reduce data mining costs, these data can be classified or
clustered before data mining. Based on the outcomes of clustering
or classification, data sources relevant to a given data source min-
ing application can be identified easily. This step is referred as data
source selection.
3.1. Data source classification
Data source classification [52] is a preprocess for multiple data
source mining, which is to classify given data sources by compar-
ing their features and measuring the relevance between different
data sources using a classifier [53]. Normally, the performance of
the classifier is measured in three aspects: (i) accuracy, (ii) com-
putational complexity, and (iii) the simplicity of the model descrip-
tions.
Wu et al. [53] propose an application-independent multiple
data source classification method, which is called BestClassifica-
tion. This method has two phases. In the first phase, it divides
all data sources by using the proposed GreedyClass algorithm. The
data sources are divided into several smaller sets according to their
similarity/distances (the similarity of two databases is measure by
the number of the common itemsets that both database have di-
vided by the total itemsets of both datasets) when a threshold is
assigned. In the second phase, it finds a best classification for these
data sources by using the proposed BestClassification algorithm.
If there is no such the best classification for these data sources,
a trivial best classification is produced. The best classification in
this paper means that all data sources can be classified properly.
However, since the time complexity of GreedyClass is O(n4 ), when
n (the number of databases) is very large, the time complexity
will become very large. At the same time, BestClassification does
not always obtain the best classification from the multiple data
sources.
An improved data classification algorithm for multiple data
source classification is proposed by Li et al. [54]. The new algo-
rithm, named as CompleteClass, has less time complexity compar-
ing with GreedyClass in [53] since CompleteClass sets a smaller
similarity threshold and decreases the number of data sources in
a class. The complexity of CompleteClass is O(n3 ), where n is the
number of databases. BestCompleteClass, an improve version of
BestClassification, can always achieve the complete classification
(i.e., the best classification) for multiple data sources. Theoretical
analysis and experimental results verify the effectiveness of the al-
gorithm. However, the running time of this method is still long
when dealing with a large amount of data.
Based on these two classification algorithms, a new multiple
data source classification method is presented by Dingrong et al.
 R. Wang et al. / Pattern Recognition Letters 109 (2018) 120–128
[55], which is called application-independent database classifica-
tion. It can achieve higher intra-class cohesion and lower inter-
class coupling compared with BestClassification and BestComplete-
Class. In order to search for the best complete classification, there
are three phases in this new algorithm. In the first phase, it
computes the similarity between each data source. In the second
phase, it chooses several databases from all data sources and clas-
sifies them into two classes based on their similarities. In the final
phase, it classifies other databases from all data sources into the
given classes in the second phase. Then it will inspect whether the
classification satisfy the given condition, if it does then end the
classification, otherwise recursive calls for further classification. To
prove the efficiency of the new algorithm, Dingrong et al. com-
pares the new method with the methods in [54] and [53]. The ex-
perimental results show the new method uses less time cost to
achieve the best complete classification for any kind of multiple
data sources.
3.2. Data source clustering
Data source clustering is another data preprocess for multiple
data source mining. Different from data source classification, it is
a kind of unguided learning. In other words, data clustering is
to cluster data according to the similarity between data without
knowing the classes they belong to in advance.
A simple method for multiple data source clustering to achieve
high cohesion and low coupling is proposed by Tang and Zhang
[56]. Similar to the proposed method in [53], it clusters data
sources based on the similarity between each data source. The
proposed method firstly constructs a multi-objective optimization
problem, then uses a hierarchical clustering algorithm, which clus-
ters multiple data sources based on the similarity between each
data sources. The hierarchical clustering algorithm can avoid in-
stability and reduce time complexity, to find the optimal cluster
structure. The multi-objective optimization problem is to find a set
of decision variables, which satisfy constraints as well as all objec-
tive functions and function values for each target (Pareto optimal
solutions), and provide the decision variables to a decision maker.
The best cluster is evaluated by the cophenetic correlation coeffi-
cient. Comparing with the BestClassification algorithm, the experi-
mental results show that this clustering method has stronger sta-
bility, lower time complexity and stronger generalization ability.
For big data, Xin et al. [57] provide two novel algorithms for
multiple data source clustering: a multi-source cross-domain clas-
sification (MSCC) algorithm, which is based on the existing max-
imum consistency function model and Newton gradient descent
method, and the MSCC Dual coordinate descent method, which is
suitable to deal with large dataset.
The MSCC algorithm is proposed to implement efficient cross-
domain classification of target domains using a logistic regression
model and a proposed consensus measure. After that, in order to
use MSCC to deal with big data in the field of multiple data source
mining, a MSCC’s fast version MSCC-CDdual is derived and theo-
retically analyzed. MSCC-CDdual is based on the algorithm CDdual
(Dual coordinate descent method) [58] as the recent advance on
large-scale logistic regression which can achieve a fast speed, high
classification accuracy and good domain adaption for multiple big
data source mining.
In [57], it proposed a new coherence function model, which
makes it easy to evolve the MSCC algorithm into a fast algorithm
for large samples. MSCC can reduce the error rate of the classifier
trained from a single source domain and improve the accuracy of
comprehensive classification. A new fast algorithm MSCC-CDdual
shows its operating speed advantage from the perspective of high-
dimensional data in [57]. Unlike the existing cross-domain algo-
rithms, MSCC is oriented to multiple data source domain learn-
125
Table 1
Multiple source data fusion methods in different fusion levels.
Pixel-level Feature-level Decision-level
Algebraic method Bayesian Knowledge-based fusion
HIS transform Dempster_shafer Dempster_shafer
High – pass filtering Entropy method Fuzzy set theory
Regression model Weighted average Reliability theory
Best variable substitution Neural network Bayesian
Kalman filter Clustering Neural network
Wavelet transform Voting Logical template
ing rather than single data source domains. Thus, it has a higher
data mining accuracy and data mining speed. However, for multi-
ple high-dimensional big data sources, the classifying and mining
efficiency of this method still need to be improved.
4. Key methods for data source fusion
Multiple data source fusion is a process of combining data from
multiple data sources to achieve higher accuracy and more specific
ratiocinations. Multiple data source fusion has been used in many
areas, including image processing (i.e., multi-sensor data fusion).
Multi-sensor data fusion is an intelligent synthesis process of
remote sensing image data from multiple sources, and to generate
accurate, complete and reliable estimation and judgment [59]. The
process is to collect the information from different sensors firstly,
and then to eliminate the data redundancy and data conflictions
that may exist among the multi-sensor information. Multi-sensor
data fusion can improve the timeliness and reliability of remote
sensing information. This is very useful for decision making, plan-
ning, and verification [60]. The main application areas of data fu-
sion include multiple data source image composite [61], robot and
intelligent instrument system [62], battlefield and unmanned air-
craft [63], image analysis and understanding, target detection and
tracking [64], and automatic target recognition [65].
There are three types of data fusion methods for remote sens-
ing images: (i) at the pixel level, (ii) at the feature level, and (iii) at
decision level. The level of fusion is in an order from low to high.
The process of multi-sensor data fusion has four steps. In the first
step, remote sensing image data from multiple data sources. The
second step is to do feature extraction. Pixel, feature or decision
level fusion is used in the third step. The last step it to do fusion
attribute description. Table 1 lists several multiple source data fu-
sion methods in different fusion levels.
4.1. Pixel-level multi-sensor data fusion
Pixel-level fusion is a low level of data fusion. The advantages
of pixel-level fusion are to retain as much information as possi-
ble. This fusion approach has the highest precision. The main pixel-
level fusion methods include Principal Component Transformation
(PCT) [66], HIS transform [68], and wavelet transform [68].
Principal Component Transformation (PCT), which is also called
Karhunen–Loeve Transform (K–L transform), is a multidimensional
linear transformation based on statistical features of an image. It
has the effect of variance information enrichment and data com-
pression, which can more accurately reveal the remote sensing in-
formation within the multi-band data structure [66]. The process
of remote sensing data fusion by using PCT has four steps [67].
The first step uses the input of multi-band remote sensing data for
spatial registration and principal component transformation. The
second step uses the results of spatial registration and principal
component transformation for histogram matching. The third step
uses the results of histogram matching to replace the fusion pro-
cess components in the first principal multi-source remote sensing
126 R. Wang et al. / Pattern Recognition Letters 109 (2018) 120–128
data. The last step generates a multi-band fused image with high
spatial resolution by using principal component inverse conversion.
Due to the physical constraints between spatial and spectral
resolutions, poor-resolution multispectral (MS) data is desirable.
Thus, a high-resolution panchromatic observation needs to be pro-
cessed before multispectral image fusion. Two general and formal
solutions, which are based on undecimated discrete wavelet trans-
form and generalized Laplacian pyramid, respectively, are proposed
by Aiazzi et al. [69]. Undecimated discrete wavelet transform is an
octave band pass representation, which is to omit all decimators
and upsample the wavelet filter bank from a conventional discrete
wavelet transform. Generalized Laplacian pyramid is another over-
sampled structure, which is obtained by recursively subtracting an
expanded decimated lowpass version from an image. The proposed
solutions based on these two methods above are to selectively per-
form spatial-frequency spectrum substitution from one image to
another. Different from other multiscale fusion methods, the ad-
vantages of the proposed methods include: (i) Context dependency
in the proposed solutions is achieved by thresholding local corre-
lation coefficient between the images to be merged. This can avoid
some spatial details which are not likely occurring to be injected
into the target image. (ii) The proposed solutions use uncritically
subsampling to avoid possible impairments in the fused images
due to missing cancellation of aliasing terms. The results presented
and discussed on the SPOT data [71] show its effectiveness.
In recent years, wavelet decomposition has become an attrac-
tive wavelet-based fusion technique to fuse multisensor images.
Normally, the input images are decomposed using an orthogonal
wavelet to extract features, which are combined through an ap-
propriate fusion rules. Then the fused image is reconstructed by
applying inverse wavelet transform. A nonorthogonal (or redun-
dant) wavelet is proposed in [70] as an alternative method for fea-
ture extraction. Compared to the normal orthogonal wavelet de-
composition, the redundant wavelet decomposition is superior to
the standard orthogonal wavelet decomposition, regardless of the
fusion rules.
After that, a new alternative method for multispectral and
panchromatic image fusion based on a high-pass filtering proce-
dure is proposed by González-Audícana et al. [68]. Multi-resolution
wavelet decomposition is used to perform detail feature extraction.
In this paper, Intensity Hue-Saturation (IHS) and Principal Compo-
nent Analysis (PCA) are used to inject the spatial details of a full-
color image into a multispectral one. A higher quality image can
be obtained due to the selective inclusion of the spatial details of
the full-color image, which is missed in the multispectral image.
The method proposed improves the quality of synthetic images
by standard IHS, PCA and standard wavelet based fusion methods.
For the two proposed fusion methods in [68], better results are
obtained when an undecimated algorithm is used to perform the
multi-resolution wavelet decomposition.
4.2. Feature-level multi-sensor data fusion
Feature-level fusion is a medium level of data fusion, which
includes three steps. The first step extracts features from all re-
mote sensing image data, the extracted features should be a suffi-
cient representation of the original information or statistically suf-
ficient. According to the feature information, all data from different
sources is classified, clustered and synthesized to generate eigen-
vectors in the second step. In the third step, some feature-level
fusion methods are used to fuse these generated eigenvectors to
make a feature description based on fusion eigenvectors.
A feature-level image fusion method based on segmented re-
gions and neural networks is proposed by Wang et al. [72] to pro-
cess the images from CCD (Charge-coupled Device) devices. Firstly,
a source image is divided into a set of common areas as a prepro-
cessing of the entire image. After that, the proposed method will
select the corresponding segmentation areas from the source im-
age, respectively, and extracts features which represent the clarity
of the corresponding regions. Then the extracted features will be
feed into a neural network to determine a clear region to recon-
struct the final fusion image. The experimental results show that
the effectiveness of this new fusion method was superior to previ-
ous methods.
A new scalable feature-level sensor fusion architecture is pro-
posed by Kaempchen et al. [73], which combines the data from a
multi-layer laser scanner with a monocular video to increase the
functionality of driver assistance systems. In order to enhance the
precision of object tracking, the proposed method established a
new geometric object model for a large diversity of object shapes
recognition. Besides, a new free form object tracking approach is
used for a precise velocity estimation of trucks on highways. The
proposed method is designed to maximize synergetic by combin-
ing feature-level measurement features and keeping the fusion ar-
chitecture as general as possible. Experimental results in an ac-
tual traffic scene show that the proposed method outperforms the
other known reference algorithms.
Canonical Correlation Analysis (CCA) is a dimensionality reduc-
tion algorithm commonly used in image processing. When extract-
ing various features from an image, each feature can form a linear
space, and CCA can be used to analyze the correlation between
these spaces. Based on CCA, a new feature extraction method
based on feature fusion is proposed by Sun et al. [74]. It is the
first time to apply CCA to feature fusion and image recognition.
The proposed method extracts two groups of feature vectors with
the same pattern firstly, then uses the correlation feature of two
groups of feature to form effective discriminant vector for recog-
nition. The proposed method not only is suitable for information
fusion, but also eliminates the redundant information within the
features, which offers a new way of applying two groups of fea-
ture fusion to discrimination. Experimental results on the Concor-
dia University CENPARMI database of handwritten Arabic numer-
als and Yale face database show that the recognition rate is much
higher than the use of a single feature or existing fusion algo-
rithms.
4.3. Decision-level multi-sensor data fusion
Decision-level fusion is cognitive-based fusion, which is the
highest level of data fusion and the highest level of abstraction.
Aiming at the specific requirements of the proposed problem,
decision-level fusion utilizes several sub-decisions or features to
yield a final or higher decision.
Decision-level fusion includes two phases. The first phase is to
make feature description for each data. The second phase fuses
the results in the last phase to get the fusion feature description
for the target or environment. The advantages of the decision-level
fusion include: strong fault tolerance, good openness, short pro-
cessing time, low data requirement and strong analytical ability.
Among the three fusion levels, the computational complexity of
decision-level fusion is the smallest. However, due to the higher
requirements for feature extraction, the cost of the decision-level
fusion is higher than the other two levels of fusion. In some cases,
decision-level fusion has a strong dependence on the other two
levels fusion.
In practice, a new method for automatic ovarian tumor classi-
fication based on the decision level fusion is proposed by Khazen-
dar et al. [75]. Firstly, the proposed method extracts two different
types of features (i.e., histograms and local binary patterns) from
ultrasound images of an ovary. Then ovarian tumor is classified ac-
cording to the features of an ultrasound image using Support Vec-
tor Machine (SVM). At this stage, the proposed method uses the
 R. Wang et al. / Pattern Recognition Letters 109 (2018) 120–128
new decision-level fusion method, which classifies SVM-based de-
cision scores into a confidence measure to help the final diagnosis
decision making. Experimental results show that such confidence-
based prediction outcomes are much closer to the reality in diag-
nosis of ovarian cancers.
The Facial Action Coding System (FACS) is the taxonomy of hu-
man facial expressions designed to facilitate human annotation of
facial expressions. The atomic facial muscle actions in this system
are called Action Units (AUs). Based on a new region-based face
representation and a mid-level region-specific decision layer, a new
method for AU detection is proposed in [76]. The proposed method
uses domain knowledge (i.e., expert knowledge of how FACS is de-
fined) regarding AU-specific facial muscle contractions to define
a set of facial regions covering the entire face. Other approaches,
however, typically represent a face as a regular grid based on
the locations on a face only (holistic representation), which may
not utilize the full facial appearance. The proposed method then
uses an AU-specific weighted sum model as a decision-level fusion
layer for region-specific probabilistic information combination. It
allows each classifier to learn the typical appearance changes for
a specific face part and reduces the dimensionality of the problem.
Experimental results based on the DISFA [77] and GEMEP-FERA
[78] datasets show superior performance of both domain-specific
region definition and decision-level fusion.
More recent, a novel multi-sensor fusion framework is proposed
by Kumar et al. [79] using Coupled Hidden Markov Model (CHMM)
for Sign Language Recognition (SLR). This framework calibrates two
different sensors (named Leap Motion [80] and Kinect [81]) with
different depth information and frames per second (fps) to form a
multi-sensor data framework for capturing dynamic sign gestures.
Leap Motion is kept below a signer’s hand to sense the movements
of signer’s hand and fingers in horizontal. Kinect is kept in front
of the signer to acquire the movements of signer’s hand and fin-
gers in vertical. Experimental results prove that new CHMM fusion
framework is better than the single sensor-based SLR camera sys-
tem.
5. Conclusion and future work
With the continuous development of data mining technology,
research in multiple data sources mining is becoming more im-
perative and important. It has a wide range of applications in the
fields such as robotics, automation and intelligent system design.
This has been and will continue to be a growing interest in the
research community to develop more advanced data mining meth-
ods and architectures. This paper critically reviewed many useful
methods to mine meaningful information and discover new knowl-
edge from multiple data sources: (i) pattern analysis, (ii) multiple
data source classification, (iii) multiple data source clustering, and
(iv) multiple data source fusion. There are still several challenges
in these three effective approaches, which need further research.
Acknowledgments
This work is in part supported by the Marsden Fund,
New Zealand, the National Natural Science Foundation of China
(U1435214, U1609215, 51507084, 61363037), the Natural Sci-
ence Foundation of Zhejiang Province, China (LY18F020 0 08), the
Youth Innovation Research Team of Sichuan Province, China
(2015TD0020), CDUT (KYTD201404), the Primary R&D Plan of
Jiangsu Province, China (BE2015213), the China Postdoctoral Sci-
ence Foundation (2016M591890), the National Social Science Foun-
dation of China (17BJY033), and the Chinese Scholarship Council
PhD scholarships.
127
References
[1] X. Zhu, L. Zhang, Z. Huang, A sparse embedding and least variance encoding
approach to hashing, IEEE Trans. Image Process. 23 (9) (2014) 3737–3750.
[2] X. Zhu, Z. Huang, H. Cheng, J. Cui, H.T. Shen, Sparse hashing for fast multime-
dia search, ACM Trans. Inf. Syst. (TOIS) 31 (2) (2013) 9.
[3] A. Adhikari, P. Ramachandrarao, W. Pedrycz, Mining multiple large databases,
in: Developing Multi-Database Mining Applications, 2010, pp. 37–50.
[4] http://www.theaustralian.com.au/national-affairs/treasury/
australian- taxation- office- turns- to- facebook- to- catch- cheats/news- story/
6d4bca2d0223dcb3924c061fa28bc3f9.
[5] S. Zhang, C. Zhang, Q. Yang, Data preparation for data mining, Appl. Artif. Intell.
17 (5–6) (2003) 375–381.
[6] S. Zhang, Z. Qin, C.X. Ling, S. Sheng, “Missing is useful”: missing values in
cost-sensitive decision trees, IEEE Trans. Knowl. Data Eng. 17 (12) (2005)
1689–1693.
[7] R. Hu, X. Zhu, D. Cheng, W. He, Y. Yan, J. Song, S. Zhang, Graph self-represen-
tation method for unsupervised feature selection, Neurocomputing 220 (2017)
130–137.
[8] X. Zhu, H.I. Suk, S.W. Lee, D. Shen, Subspace regularized sparse multitask
learning for multiclass neurodegenerative disease identification, IEEE Trans.
Biomed. Eng. 63 (3) (2016) 607–618.
[9] X. Zhu, X. Li, S. Zhang, Block-row sparse multiview multilabel learning for im-
age classification, IEEE Trans. Cybern. 46 (2) (2016) 450–461.
[10] M.W. Bright, A.R. Hurson, S.H. Pakzad, A taxonomy and current issues in mul-
tidatabase systems, Computer 25 (3) (1992) 50–60.
[11] S. Zhang, X. Li, M. Zong, X. Zhu, R. Wang, Efficient kNN classification with dif-
ferent numbers of nearest neighbours, IEEE Trans. Neural Networks Learn. Syst.
(2017), doi:10.1109/TNNLS.2017.2673241.
[12] S. Zhang, X. Li, M. Zong, X. Zhu, D. Cheng, Learning k for knn classification,
ACM Trans. Intell. Syst. Technol. (TIST) 8 (3) (2017) 43.
[13] X. Zhu, X. Li, S. Zhang, C. Ju, X. Wu, Robust joint graph sparse coding for un-
supervised spectral feature selection, IEEE Trans. Neural Networks Learn. Syst.
26 (6) (2016) 1263–1275.
[14] S. Zhang, Z. Jin, X. Zhu, Missing data imputation by utilizing information
within incomplete instances, J. Syst. Softw. 84 (3) (2011) 452–459.
[15] A. Turinsky, R. Grossman, A framework for finding distributed data mining
strategies that are intermediate between centralized strategies and in-place
strategies, in: Proceedings of Workshop on Distributed and Parallel Knowledge
Discovery at KDD-20 0 0, 20 0 0, pp. 1–7.
[16] X. Wu, C. Zhang, S. Zhang, Database classification for multi-database mining,
Inf. Syst. 30 (1) (2005) 71–88.
[17] S. Zhang, C. Zhang, X. Wu, Data mining and multi-database mining, in: Knowl-
edge Discovery in Multiple Databases, 2004, pp. 27–61.
[18] S. Zhang, M.J. Zaki, Mining multiple data sources: local pattern analysis, Data
Mining Knowl. Discov. 12 (2) (2006) 121–125.
[19] X. Wu, S. Zhang, Synthesizing high-frequency rules from different data sources,
IEEE Trans. Knowl. Data Eng. 15 (2) (2003) 353–367.
[20] S. Zhang, X. You, Z. Jin, X. Wu, Mining globally interesting patterns from
multiple databases using kernel estimation, Expert Syst. Appl. 36 (8) (2009)
10863–10869.
[21] X. Yan, C. Zhang, S. Zhang, Toward databases mining: preprocessing collected
data, Appl. Artif. Intell. 17 (5–6) (2003) 545–561.
[22] S. Zhang, Q. Chen, Q. Yang, Acquiring knowledge from inconsistent data
sources through weighting, Data Knowl. Eng. 69 (8) (2010) 779–799.
[23] S. Zhang, M. Zong, K. Sun, Y. Liu, D. Cheng, Efficient kNN algorithm based on
graph sparse reconstruction, in: International Conference on Advanced Data
Mining and Applications, 2014, pp. 356–369.
[24] X. Zhu, S. Zhang, Z. Jin, Z. Zhang, Z. Xu, Missing value estimation for mixed-at-
tribute data sets, IEEE Trans. Knowl. Data Eng. 23 (1) (2011) 110–121.
[25] S. Zhang, X. Wu, C. Zhang, Multi-database mining, IEEE Comput. Intell. Bull. 2
(1) (2003) 5–13.
[26] Y. Chen, F. Li, J. Fan, Mining association rules in big data with NGEP, Cluster
Comput. 18 (2) (2015) 577–585.
[27] R. Agrawal, R. Srikant, Fast algorithms for mining association rules, in: 20th In-
ternational Conference Very Large Data Bases, VLDB, 1215, 1994, pp. 487–499.
[28] J. Han, J. Pei, Y. Yin, R. Mao, Mining frequent patterns without candidate gen-
eration: a frequent-pattern tree approach, Data Mining Knowl. Discov. 8 (1)
(2004) 53–87.
[29] J. Yan, N. Liu, Q. Yang, B. Zhang, Q. Cheng, Z. Chen, Mining adaptive ratio rules
from distributed data sources, Data Mining Knowl. Discov. 12 (2–3) (2006)
249–273.
[30] Z. Abdullah, T. Herawan, N. Ahmad, M.M. Deris, Mining significant association
rules from educational data using critical relative support approach, Proc. Soc.
Behav. Sci. 28 (2011) 97–101.
[31] J. Sahoo, A.K. Das, A. Goswami, An efficient approach for mining association
rules from high utility itemsets, Expert Syst. Appl. 42 (13) (2015) 5754–5778.
[32] R. Agrawal, R. Srikant, Mining sequential patterns, in: Proceedings of the
Eleventh International Conference on Data Engineering, 1995, pp. 3–14.
[33] J. Han, J. Pei, B. Mortazavi-Asl, H. Pinto, Q. Chen, U. Dayal, M.C. Hsu, Prefixspan:
mining sequential patterns efficiently by prefix-projected pattern growth, in:
Proceedings of the 17th International Conference on Data Engineering, 2001,
pp. 215–224.
[34] H.C. Kum, J.H. Chang, W. Wang, Sequential pattern mining in multi-databases
via multiple alignment, Data Mining Knowl. Discov. 12 (2-3) (2006) 151–180.
128 R. Wang et al. / Pattern Recognition Letters 109 (2018) 120–128
[35] P. Fournier-Viger, A. Gomariz, M. Campos, R. Thomas, Fast vertical mining of
sequential patterns using co-occurrence information, in: Pacific-Asia Confer-
ence on Knowledge Discovery and Data Mining, 2014, pp. 40–52.
[36] M.J. Zaki, SPADE: an efficient algorithm for mining frequent sequences, Mach.
Learn. 42 (1) (2001) 31–60.
[37] J. Ayres, J. Flannick, J. Gehrke, T. Yiu, Sequential pattern mining using a bitmap
representation, in: Proceedings of the Eighth ACM SIGKDD International Con-
ference on Knowledge Discovery and Data Mining, 2002, pp. 429–435.
[38] A. Gomariz, M. Campos, R. Marin, B. Goethals, ClaSP: an efficient algorithm for
mining frequent closed sequences, in: Pacific-Asia Conference on Knowledge
Discovery and Data Mining, 2013, pp. 50–61.
[39] Y. Wu, L. Wang, J. Ren, W. Ding, X. Wu, Mining sequential patterns with peri-
odic wildcard gaps, Appl. Intell. 41 (1) (2014) 99–116.
[40] X. Yan, J. Han, gspan: graph-based substructure pattern mining, in: Proceed-
ings of IEEE International Conference on Data Mining, 2002, pp. 721–724.
[41] R. Suganthi, P. Kamalakannan, Exceptional patterns with clustering items in
multiple databases, Indian J. Sci. Technol. 8 (31) (2015) 1–10.
[42] C. Zhang, M. Liu, W. Nie, S. Zhang, Identifying global exceptional patterns in
multi-database mining, IEEE Intell. Inf. Bull. 3 (1) (2004) 19–24.
[43] T. Ramkumar, R. Srinivasan, Modified algorithms for synthesizing high-fre-
quency rules from different data sources, Knowl. Inf. Syst. 17 (3) (2008)
313–334.
[44] S. Khiat, H. Belbachir, S.A. Rahal, Probabilistic models for local patterns analy-
sis, University of Sciences and Technology of Oran, 2015.
[45] A. Adhikari, P.R. Rao, Synthesizing heavy association rules from different real
data sources, Pattern Recognit. Lett. 29 (1) (2008) 59–71.
[46] D.L. Donoho, Compressed sensing, IEEE Trans. Inf. Theory 52 (4) (2006)
1289–1306.
[47] B. Yang, S. Li, Multifocus image fusion and restoration with sparse representa-
tion, IEEE Trans. Instrum. Meas. 59 (4) (2010) 884–892.
[48] J. Tao, S. Wen, W. Hu, Multi-source adaptation learning with global and local
regularization by exploiting joint kernel sparse representation, Knowl. Based
Syst. 98 (2016) 76–94.
[49] R. Porter, V. Tadic, A. Achim, Sparse Bayesian learning for non-Gaussian
sources, Digit. Signal Process. 45 (2015) 2–12.
[50] A. Adhikari, P. Ramachandrarao, B. Prasad, J. Adhikari, Mining multiple large
data sources, Int. Arab J. Inf. Technol. 7 (3) (2010) 241–249.
[51] A. Adhikari, L.C. Jain, S. Ramanna, Analysing effect of database grouping on
multi-database mining, IEEE Intell. Inf. Bull. 12 (1) (2011) 25–32.
[52] X. Zhu, Q. Xie, Y. Zhu, X. Liu, S. Zhang, Multi-view multi-sparsity kernel re-
construction for multi-class image classification, Neurocomputing 169 (2015)
43–49.
[53] X. Wu, C. Zhang, S. Zhang, Database classification for multi-database mining,
Inf. Syst. 30 (1) (2005) 71–88.
[54] H. Li, X. Hu, Y. Zhang, An improved database classification algorithm for mul-
ti-database mining, in: International Workshop on Frontiers in Algorithmics,
2009, pp. 346–357.
[55] Y.U.A.N. Dingrong, F.U. Huiwen, L.I. Zhixin, W.U. Hongyi, An application-inde-
pendent database classification method based on high cohesion and low cou-
pling, J. Inf. Comput. Sci. 9 (15) (2012) 4337–4344.
[56] H. Tang, M. Zhang, A simple methodology for database clustering, in: Proceed-
ings of Science (CENet), 2015, pp. 1–7.
[57] G. Xin, S. Wang, M. Xu, A new cross-multidomain classification algorithm and
its fast version for large datasets, Acta Autom. Sin. 40 (3) (2014) 531–547.
[58] C.J. Hsieh, K.W. Chang, C.J. Lin, S.S. Keerthi, S. Sundararajan, A dual coordinate
descent method for large-scale linear SVM, in: Proceedings of the 25th Inter-
national Conference on Machine Learning, 2008, pp. 408–415.
[59] B. Khaleghi, A. Khamis, F.O. Karray, S.N. Razavi, Multisensor data fusion: a re-
view of the state-of-the-art, Inf. Fusion 14 (1) (2013) 28–44.
[60] R.C. Luo, M.H. Lin, R.S. Scherp, Dynamic multi-sensor data fusion system for
intelligent robots, IEEE J. Robot. Autom. 4 (4) (1988) 386–396.
[61] Y. Zhang, S. Prasad, Locality preserving composite kernel feature extraction for
multi-source geospatial image analysis, IEEE J. Selected Top. Appl. Earth Ob-
serv. Remote Sens. 8 (3) (2015) 1385–1392.
[62] R.C. Luo, M.G. Kay, Multisensor integration and fusion in intelligent systems,
IEEE Trans. Syst. Man Cybern. 19 (5) (1989) 901–931.
[63] M.P. Đurišić, Z. Tafa, G. Dimić, V. Milutinović, A survey of military applications
of wireless sensor networks, in: 2012 Mediterranean Conference on Embedded
Computing (MECO), 2012, pp. 196–199.
[64] Y. Bar-Shalom, P.K. Willett, X. Tian, Tracking and Data Fusion, YBS Publishing,
2011.
[65] J. Llinas, D.L. Hall, An introduction to multi-sensor data fusion, in: Proceed-
ings of the 1998 IEEE International Symposium on Circuits and Systems, 1998.
ISCAS’98, 6, 1998, pp. 537–540.
[66] A. Dang, X. Wang, X. Chen, in: ERDAS IMAGINE Remote Sensing Image Pro-
cessing Method, Tsinghua University Press, 2003, pp. 1–6.
[67] G. Wei, S. Yang, Multisource RS data fusion based on principal component
analysis, J. Lanzhou Jiaotong Univ. (Nat. Sci.) 24 (3) (2005) 57–60.
[68] M. González-Audícana, J.L. Saleta, R.G. Catalán, R. García, Fusion of multispec-
tral and panchromatic images using improved IHS and PCA mergers based
on wavelet decomposition, IEEE Trans. Geosci. Remote Sens. 42 (6) (2004)
1291–1299.
[69] B. Aiazzi, L. Alparone, S. Baronti, A. Garzelli, Context-driven fusion of high spa-
tial and spectral resolution images based on oversampled multiresolution anal-
ysis, IEEE Trans. Geosci. Remote Sens. 40 (10) (2002) 2300–2312.
[70] Y. Chibani, A. Houacine, Redundant versus orthogonal wavelet decomposition
for multisensor image fusion, Pattern Recognit. 36 (4) (2003) 879–887.
[71] P. Terretaz, Comparison of different methods to merge SPOT P and XS data:
evaluation in an urban area, in: Proceedings of the 17th Symposium of EARSeL,
Future Trends in Remote Sensing, Lyngby, Denmark, 1997, pp. 17–20.
[72] R. Wang, F. Bu, H. Jin, L. Li, A feature-level image fusion algorithm based
on neural networks, in: Proceedings of the First International Conference on
Bioinformatics and Biomedical Engineering, 2007, pp. 821–824.
[73] N. Kaempchen, M. Buehler, K. Dietmayer, Feature-level fusion for free-form ob-
ject tracking using laserscanner and video, in: Proceedings IEEE Intelligent Ve-
hicles Symposium, 2005, pp. 453–458.
[74] Q.S. Sun, S.G. Zeng, Y. Liu, P.A. Heng, D.S. Xia, A new method of feature fu-
sion and its application in image recognition, Pattern Recognit. 38 (12) (2005)
2437–2448.
[75] S. Khazendar, H. Al-Assam, et al., Automated classification of static ultrasound
images of ovarian tumours based on decision level fusion, in: Computer Sci-
ence and Electronic Engineering Conference (CEEC), 2014, pp. 148–153.
[76] B. Jiang, B. Martinez, M.F. Valstar, M. Pantic, Decision level fusion of domain
specific regions for facial action recognition, in: 2014 22nd International Con-
ference on Pattern Recognition (ICPR), 2014, pp. 1776–1781.
[77] S.M. Mavadati, M.H. Mahoor, K. Bartlett, P. Trinh, Automatic detection of non–
posed facial action units, in: 2012 19th IEEE International Conference on Image
Processing (ICIP), 2012, pp. 1817–1820.
[78] M.F. Valstar, M. Mehu, B. Jiang, M. Pantic, K. Scherer, Meta-analysis of the first
facial expression recognition challenge, IEEE Trans. Syst. Man Cybern. Part B
(Cybern.) 42 (4) (2012) 966–979.
[79] P. Kumar, H. Gauba, P.P. Roy, D.P. Dogra, Coupled hmm-based multi-sensor data
fusion for sign language recognition, Pattern Recognit. Lett. 86 (2017) 1–8.
[80] https://www.leapmotion.com/.
[81] https://developer.microsoft.com/en-us/windows/kinect.